CN103297773B - Based on the method for encoding images of JND model - Google Patents

Abstract

Patent of the present invention relates to a kind of method for encoding images based on JND model, complete after the dct transform of image, the module being made up of quantification, inverse quantization reconstruct, JND threshold estimation and four parts of enhancement quantized device further increases quantization step, remove visual redundancy, in the situation that keeping visual quality of images constant, reduce compression bit rate, improve code efficiency. The method can accurately be followed the tracks of according to user the quality of compressed image to the needs of compression bit rate, provide accurate JND threshold value to design and quantize enhancing value. In addition, the improved JND threshold estimation of the present invention method makes coding side not need to increase extra overhead bit, and the code stream that the encoder of design produces keeps the compatibility with JPEG coding standard.

Description

Based on the method for encoding images of JND model

Technical field

Patent of the present invention relates to a kind of based on JND (JustNoticeableDifference, JND) modelMethod for encoding images, relates in particular to a kind of method for encoding images based on visual characteristic.

Background technology

In image is processed and transmitted, Joint Photographic Experts Group is a kind of widely used image storage and transformat,It is International Organization for Standardization subordinate's the JPEG(JointPhotographic of JPEGExpertsGroup) JPEG of formulating.

Joint Photographic Experts Group coding method as shown in Figure 1, mainly comprised image input, dct transform, quantification withAnd entropy coding module. Although this coding techniques has obtained the extensive accreditation of industry, it is based upon Shannon letterOn breath opinion basis, taking classical sets opinion as instrument, describe information source with probability statistics model, it compresses thoughtStill mainly rest on the aspect of Digital Signal Processing, mainly start with from removing data redundancy aspect, less consideration is lookedRedundancy properties in feel, the raising of code efficiency mainly depends on the skill that significantly increases to cost with computational complexityThe fine setting of art details. Along with the progressively development of Image Coding standard, promote compression effect to improve computation complexityThe room for improvement of rate is more and more less, and this has become the root making a breakthrough in restriction Vision information processing fieldThis bottleneck, the urgently innovation in theory and method.

The basic goal of coding is can provide high-quality visual effect in reducing code check, a kind of directly perceivedEfficient coding thinking should be from the mankind's visual characteristic, under the condition that keeps visual quality as much as possibleReduce code check. In recent years, along with to human vision research gradually deeply, it is found that the angle from human visionDegree sets out, and will have very large compression stroke. JND model is to set up perception by exploring all kinds of visual signaturesThe threshold value of error, distinguishes that can perceive and the not perceived signal of people, and then it is superfluous to remove visually-perceptibleMore than. When image change value is during lower than threshold value, human eye will can not perceive this variation. In order to overcome current imageThe shortcoming of coding standard, patent of the present invention in conjunction with JND algorithm, through adjustment after JND models coupling to figureIn picture cataloged procedure, improve code efficiency.

Summary of the invention

Patent of the present invention proposes a kind of method for encoding images based on JND model, the JND after adjustingModels coupling, in quantizing process, increases quantify strength, removes more visual redundancy, in equal vision matterUnder amount, improve code efficiency, reduced compression bit rate, and the compatibility of maintenance and JPEG coding standard.

A kind of method for encoding images based on JND model of the present invention, is characterized in that: image input module is provided,Dct transform device, quantizer, inverse DCT, DCT inverse transform block, JND computing module, enhancement quantized deviceAnd entropy coding module is undertaken by following flow process:

Raw image files is read by image input module and image is carried out to piecemeal;

Dct transform device carries out DCT direct transform and exports DCT coefficient each 8x8 piecemeal of input picture;

The quantization step Q that quantizer meets Joint Photographic Experts Group to DCT coefficients by using quantizes in advance, and amountChange result and export to inverse DCT;

Inverse DCT uses identical quantization step Q to complete the coefficient of inverse quantization work acquisition reconstruct;

DCT inverse transform block is carried out the image of DCT inverse transformation acquisition reconstruct to the coefficient of reconstruct;

JND computing module is to the non maximum JND threshold value of the image calculation human eye of reconstruct;

Enhancement quantized device is according to the △ Q that gains in strength of JND threshold calculations quantization step, and obtains final enhancingQuantization step Q_stepDCT coefficient is carried out to the enhancement quantized based on visual characteristic;

After entropy coding module uses the entropy encryption algorithm that meets JPEG coding standard to enhancement quantized, data are furtherCompression, realizes the compatibility with JPEG code stream.

The present invention can accurately follow the tracks of according to user the quality of compressed image to the needs of compression bit rate, provide accurateJND threshold value design and quantize enhancing value. In addition, the improved JND threshold estimation of the present invention method makes codingEnd does not need to increase extra overhead bit, and the code stream that the encoder of design produces keeps and JPEG coding standardCompatibility.

Brief description of the drawings

Fig. 1 is the method for encoding images structured flowchart of Joint Photographic Experts Group.

Fig. 2 is the method for encoding images FB(flow block) based on JND model.

Detailed description of the invention

Set forth with specific embodiment the technical scheme that patent of the present invention relates to below in conjunction with accompanying drawing.

The present invention has adopted and has quantized in advance to estimate the JND threshold value of compressed images, and enters one in conjunction with JND modelStep increases quantization step, in the situation that keeping picture quality constant, improves by removing more visual redundancyCode efficiency, the code stream of gained of the present invention keeps the compatibility with JPEG coding standard, this method for based onThe Video coding of JND model is applicable equally.

As shown in Figure 2. Raw image files is read by image input module 11 and image is carried out to piecemeal; DCTConverter 12 carries out DCT direct transform and exports DCT coefficient each 8x8 piecemeal of input picture; QuantizeThe quantization step Q that device 13 meets Joint Photographic Experts Group to DCT coefficients by using quantizes in advance, and quantized resultExport to inverse DCT; Inverse DCT 14 uses identical quantization step Q to complete inverse quantization work and obtains reconstructCoefficient; DCT inverse transform block 15 is carried out the image of DCT inverse transformation acquisition reconstruct to reconstruction coefficients; JNDComputing module 16 calculates the non maximum JND threshold value of human eye to reconstructed image; Enhancement quantized device 17 basesThe △ Q that gains in strength of JND threshold calculations quantization step, and obtain final enhancement quantized step-length Q_stepTo DCTCoefficient carries out the enhancement quantized based on visual characteristic; Entropy coding module 18 uses the entropy that meets JPEG coding standardEncryption algorithm further compresses data, realizes the compatibility with JPEG code stream.

JND computing module 16 calculates the non maximum JND threshold value of human eye to reconstructed image. Based on DCTThe JND model description of transform domain is that spatial contrast degree sensitivity function, brightness self adaptation and contrast are covered three'sLong-pending, as shown in Equation (1):

JND_DCT(i,j)=J_CSF(i,j)×A_lum×F_contrast(i,j)（1）

Wherein, (i, j) is the call number of DCT coefficient, wherein spatial contrast degree sensitivity function J_CSF(i, j) and imageContent irrelevant, decoding end can obtain in advance, therefore J_CSF(i, j) do not need to add code stream; Brightness self adaptationFactors A_lumRelevant with mean flow rate, and the DC component DC of DCT coefficient has directly embodied the water of mean flow rateFlat, formula (2) has shown the relation of JPEG encoding D CT DC coefficient and mean flow rate, wherein N=8, forThe size of image block. In order to reduce the transmission of supplementary, when DC coefficient is quantized, JND only adoptsJ_CSFThe value of (i, j) part, formula for decoder (2) calculates after mean flow rate, and recycling formula (3) is justCan obtain brightness adaptive factor A_lum, wherein DC represents DC component; N is the size of piecemeal, hereN=8；Represent mean flow rate.

$\stackrel{\‾}{I}=\frac{\mathrm{DC}}{N}+128---\left(2\right)$

$A_{\mathrm{lum}}=\begin{array}{c}\end{array}\left\{\begin{array}{cc}1+(60-\stackrel{\&OverBar;}{I})/150& \stackrel{\&OverBar;}{I}\&le;60\\ 1& 60<\stackrel{\&OverBar;}{I}<170\\ 1+(\stackrel{\&OverBar;}{I}-170)/425& \stackrel{\&OverBar;}{I}\&GreaterEqual;170\end{array}\right.---\left(3\right)$

Contrast is covered weighted factor F_contrast(i, j) is relevant to the susceptibility in different images region with human eye, whereinThe weighted value 1 of level and smooth and fringe region; Human eye is relatively little to the susceptibility of texture region low frequency coefficient, thereforeWeight coefficient is 2.25, and high frequency coefficient be weighted to 1.25. Each only needs extra increase in the time of codingAn area type information just can make decoding end obtain F_contrast(i, j). In JPEG coding standard, eachPiece all can end up with EOB (the EOB end of block character) after Huffman encoding, luminance factor in standard scaleEOB be 1010, in fact partition information can judge together with EOB, compiles according to Huffman hereinThe uniqueness principle of code has been chosen another code value 1001 and has been combined judgement with 1010, and wherein 1010 expressions are currentPiece is edge or flat site, and 1001 represent that current block is texture region. This coding method makes to compileCode end is embedded into code stream area type information in not needing to increase any bit number, has improved coding effectRate, produced simultaneously code stream keeps the compatibility with JPEG coding standard. The enhancement quantized step-length Q finally obtaining_stepAs shown in Equation (4), wherein Q is the original quantization step of JPEG coding standard; △ Q represents quantization stepGain in strength, JND_DCT(i, j) is JND threshold value.

Q_step=Q+△Q=Q+2JND_DCT(i,j)（4）。

Claims (2)

1. the method for encoding images based on JND model, is characterized in that: provide image input module, dct transform device, amountChanging device, inverse DCT, DCT inverse transform block, JND computing module, enhancement quantized device and entropy coding module enters by following flow processOK:

Raw image files is read by image input module and image is carried out to piecemeal;

Dct transform device carries out DCT direct transform and exports DCT coefficient each 8x8 piecemeal of input picture;

The quantization step Q that quantizer meets Joint Photographic Experts Group to DCT coefficients by using quantizes in advance, and quantized result is exportedGive inverse DCT;

Inverse DCT uses identical quantization step Q to complete the coefficient of inverse quantization work acquisition reconstruct;

DCT inverse transform block is carried out the image of DCT inverse transformation acquisition reconstruct to the coefficient of reconstruct;

JND computing module is to the non maximum JND threshold value of the image calculation human eye of reconstruct;

Enhancement quantized device is according to the △ Q that gains in strength of JND threshold calculations quantization step, and obtains final enhancement quantized step-lengthQ_stepDCT coefficient is carried out to the enhancement quantized based on visual characteristic;

Data after entropy coding module uses the entropy encryption algorithm that meets JPEG coding standard to enhancement quantized are further compressed,Realize the compatibility with JPEG code stream;

Described JND model description is that spatial contrast degree sensitivity function, brightness self adaptation and contrast are covered the long-pending of three, as shown in formula (1)：

JND_DCT(i,j)＝J_CSF(i,j)×A_lum×F_contrast(i,j)(1)

Wherein spatial contrast degree sensitivity function J_CSF(i, j) is irrelevant with the content of image, can obtain in advance by decoding end, thereforeJ_CSF(i, j) do not need to add code stream; Brightness adaptive factor A_lumRelevant with mean flow rate, and the DC component of DCT coefficientDC has directly embodied the level of mean flow rate, and formula (2) has shown the pass of JPEG encoding D CT DC coefficient and mean flow rateSystem, wherein N=8, is the size of image block; In order to reduce the transmission of supplementary, when DC coefficient is quantized, JND only adoptsJ_CSFThe value of (i, j) part, formula for decoder (2) calculates after mean flow rate, and recycling formula (3) just can obtain brightnessAdaptive factor A_lum；

$\stackrel{\&OverBar;}{I}=\frac{DC}{N}+128---\left(2\right)$

$A_{lum}=\left\{\begin{array}{cc}1+(60-\stackrel{\&OverBar;}{I})/150& \stackrel{\&OverBar;}{I}\&le;60\\ 1& 60<\stackrel{\&OverBar;}{I}<170\\ 1+(\stackrel{\&OverBar;}{I}-170)/425& \stackrel{\&OverBar;}{I}\&GreaterEqual;170\end{array}\right.---\left(3\right)$

Contrast is covered weighted factor F_contrast(i, j) is relevant to the susceptibility in different images region with human eye, wherein level and smooth withThe weighted value 1 of fringe region; Human eye is relatively little to the susceptibility of texture region low frequency coefficient, and therefore weight coefficient is 2.25,And high frequency coefficient be weighted to 1.25; Each only needs area type information of extra increase just can make to separate in the time of codingCode end obtains F_contrast(i,j)；

In described JPEG coding standard, each can end up with EOB after Huffman encoding, this EOB choose 1001 with1010 combine judgement, and wherein 1010 represent that current block is edge or flat site, and 1001 represent that current block isTexture region.

2. the method for encoding images based on JND model according to claim 1, is characterized in that: described enhancement quantizedStep-length Q_stepAs shown in formula (4):

Q_step＝Q+△Q＝Q+2JND_DCT(i,j)(4)。